The World’s Strongest Open Source Large-scale Model Llama 3 Released
Recently, Meta has released its latest open-source large language model, Llama 3. This model offers both 8B and 70B pre-trained and fine-tuned versions with instructions, and is designed to support a broad range of applications. It is claimed that Llama 3 is currently the most advanced open-source large language model at both 8B and 70B parameter scales. It has been integrated into Meta AI’s virtual assistant, and in the future, it will be widely applied to various Meta services.
Llama 3 employs a pure decoder transformer architecture, which has significantly improved both performance and inference efficiency. Compared to its predecessor, Llama 2, Llama 3 has enriched training data exceeding 15T, seven times the size of the previous model, and includes four times the code data. The expansion of pre-training scale and the innovative method of fine-tuning instructions have together enhanced the overall performance of Llama 3.
According to Meta’s plan, the existing 8B and 70B Llama 3 models are just a beginning. In the coming months, Meta will release a series of models with functional updates. These new models will feature advanced properties such as multimodal functions, extended context windows, and multilingual dialogue capabilities. Additionally, a Llama 3 model with over 400B parameters is currently being trained, and Meta is expected to publish a comprehensive research paper based on this large-scale model.
Several Cities in China Are Experiencing a Ground Subsidence Crisis
Ground subsidence is a serious urban problem that not only endangers the safety of urban buildings and infrastructure but also often lacks sufficient attention from the public. A study published in the journal “Science” noted that about 270 million urban residents in China are affected by ground subsidence, which is approximately one-third of the urban population. Nearly 70 million of these people live in areas where the ground subsides more than 10 millimeters per year.
The research team conducted a detailed analysis of the ground movement in 82 Chinese cities using the latest satellite data. The results showed that nearly half of the cities analyzed experienced subsidence: 16% of the cities are sinking at a rate exceeding 10 millimeters per year. The phenomenon of ground subsidence can exacerbate the impacts of climate change, especially for coastal cities such as Tianjin and Shanghai, since land subsidence may lead to rising sea levels and increased flood risks.
The study points out that groundwater extraction is the main factor causing subsidence, but geological conditions and the weight of buildings also play a role in the effect. Notably, since the cessation of groundwater extraction in the 1970s in Osaka and Tokyo, the phenomenon of ground subsidence has visibly decreased, suggesting that corresponding measures may be effective.
Researchers call on the academic community to not only measure urban subsidence but also assess its impact and help governments formulate response strategies.
Health Recent advances in the study of cognitive health in old age suggest that people who frequently use their mental faculties during middle age are at lower risk of developing mild cognitive impairment and dementia later in life. This finding was completed in collaboration by scientists from Columbia University and the Norwegian Institute of Public Health, among other institutions, and the results have been published in the prestigious Journal of Neurology. The research team integrated administrative data from Norway with the Occupational Information Network database of the U.S. Bureau of Labor, Employment, and Training Administration to assess the cognitive demands of more than 300 occupations, using the Routine Task Intensity (RTI) index to measure the level of cognitive demand. In the study, 7,003 participants aged 70 or older were divided into four groups based on their occupational RTI to assess the association between different levels of cognitive demand and the occurrence of mild cognitive impairment and dementia. After adjusting for age, gender, and education level, the results showed that groups with lower cognitive demands had a 37% increased risk of developing dementia compared to those with higher cognitive demands. This indicates a significant positive correlation between mental activity in middle age and the maintenance of cognitive abilities in later life.
Neuroscience The impact of addictive drugs on the human brain has always been a hot topic for scientific research. Although it is well known that addictive drugs and natural rewards both activate the same areas in the brain, the specific impact on brain circuits remains unclear. Recently, scientists from Mount Sinai Hospital in New York and Rockefeller University in the U.S. published new research in the journal Science, revealing how addictive drugs hijack the brain’s reward circuits. The study for the first time demonstrated how psychostimulants (such as cocaine) and opioids (such as morphine) alter the mechanisms by which specific neural cells process rewards. It focused on the differential responses of individual neurons in the nucleus accumbens to natural rewards such as food and these two drugs. The study found that certain neurons respond to both conditions, with a more potent response under drug stimulation. Long-term drug exposure disrupted the normal functions of these neurons, causing them to pursue drugs rather than natural rewards, leading to disordered responses to natural rewards during withdrawal and thereby triggering negative emotional reactions associated with withdrawal.
The scientists discovered the role of this mechanism in the destruction of neuronal processing of the natural reward system during drug abuse when researching the intracellular mTORC1 signaling pathway. They identified a gene called Rheb that is responsible for encoding the activator of this signaling pathway, a finding that provides a potential new target for treating related diseases. To better understand the molecular mechanisms behind addictive behaviors, researchers will further explore how drug abuse can hijack and alter neural circuits that should respond to natural rewards, leading to the development of addictive behaviors.
In the field of paleobiology, a new study published in the journal Scientific Reports mentions that researchers in Gujarat, India, discovered a species of giant ancient snake named Vasuki Indicus. Estimated to have lived about 47 million years ago, it may be one of the largest snakes in history. V. Indicus, estimated to be between 11 and 15 meters in length, belonged to the now-extinct madtsoiidae family of giant snakes, representing a unique lineage that originated in India. The discovery described by researchers is located at a lignite mine in the Kachchh district of Panandhro, where the fossils date back to the Middle Eocene epoch.
The article mentions 27 well-preserved vertebral bones, with some still retaining their articular connections. These vertebrae indicate that the snake had a broad, cylindrical body. Based on the size of the vertebrae, it is speculated that V. Indicus could have been between 10.9 meters to 15.2 meters long, comparable in size to the Titanoboa. However, there is a certain degree of uncertainty in this estimate. The research goes on to speculate that due to the large size of V. Indicus, it might have been a slow-moving predator that primarily ambushed its prey, similar to the modern reticulated python. Through fossil analysis, researchers have determined that V. Indicus belonged to the Boidae family, a group of snakes that lived for about 100 million years from the Late Cretaceous to the Late Pleistocene, with a distribution that covered areas including Africa, Europe, and India. The study suggests that V. Indicus may represent a type of large snake that originated on the Indian subcontinent, which likely spread to Africa via a southern route through Europe during the Eocene epoch.
